Most small and mid-size firms are not short on legal talent. They are short on the hours between the legal work: the intake follow-ups that never got sent, the medical records that arrived as a 400-page PDF and sat unreviewed for a week, the discovery deadline that nobody calendared because the scheduling order came in as an email attachment on a Friday afternoon. That gap is where cases stall and where clients decide you are unresponsive.

The phrase ai paralegal has started showing up in vendor demos as shorthand for software that absorbs some of that work. It is a useful shorthand as long as you keep the boundary clear. Nothing here practises law, exercises judgment, or takes responsibility. What an ai paralegal can do is handle the deterministic, repetitive, and document-heavy operational tasks that surround the practice of law, and do them consistently at two in the morning without being asked twice.

What actually delegates well

The tasks that transfer cleanly to software share three traits: the inputs are predictable, the output is checkable, and a mistake is visible rather than silent.

  • Intake routing and follow-up. A form submission or missed call triggers a sequence: create the matter record in Clio or MyCase, send a text within two minutes, schedule a consult, chase the no-shows on a defined cadence, and flag anything that has gone quiet for 48 hours. This is where most firms lose the most revenue and where automation is least controversial.
  • Document intake and classification. Records arrive as email attachments, fax-to-PDF, or portal downloads. Software can name them consistently, file them into the right matter folder, split combined PDFs by provider or date range, and extract dates, providers, and billing totals into a structured summary an attorney or paralegal can scan.
  • Deadline extraction and calendaring. Scheduling orders, hearing notices, and discovery responses contain dates in predictable places. Extracting them and drafting calendar entries for human confirmation is reliable. Auto-committing them without review is not.
  • First-draft routine correspondence. Records request letters, HIPAA authorisations, status updates to clients, and simple demand package cover letters follow templates. Populating them from matter data saves real time.
  • Status monitoring and nudging. Which matters have had no activity in 21 days. Which clients have not returned a signed retainer. Which medical providers have not responded to a second request. This is pure bookkeeping and it is where humans are least reliable.
  • Billing hygiene. Flagging unbilled time, matters approaching a fee cap, and entries with vague narratives before they go to a client.

Notice what these have in common. None of them require deciding what the law is or what the client should do.

A worked example: personal injury records intake

Here is a workflow that a mid-size PI firm can run today, described specifically enough to build.

A provider emails records to a dedicated intake address. The automation picks up the message, reads the subject line and body for a client name and date of loss, and matches it against open matters in Filevine. If the match confidence is low, it does not guess; it drops the item into a review queue with the two or three candidate matters listed.

Once matched, the system splits the PDF at provider boundaries and page-count breaks, then runs extraction on each segment: treating provider, date range of treatment, ICD codes present, itemised charges, and whether the file contains imaging reports or only billing. It writes a structured summary into the matter as a note and attaches the segmented files with consistent naming, something like 2024-03-14_MercyGeneral_Billing.pdf.

It then compares the extracted date range against the treatment timeline already in the matter and flags gaps. If the client reported treatment through August but records stop in June, that flag goes to the case manager, not into a demand package.

Finally, it updates a running damages worksheet with the billed amounts and marks the provider as received in the records request tracker, which stops the automated follow-up sequence to that provider.

What the software has not done: assessed whether treatment was reasonable and necessary, decided which records are prejudicial, formed a view on causation, or valued the claim. A paralegal still reads the summary. An attorney still reads the records that matter. The automation removed roughly the mechanical half of the task and made the remaining half faster to start.

Where it does not work, plainly

Be sceptical of any demo that glides past these.

Extraction from poor-quality scans is unreliable. Handwritten physician notes, faxed documents that have been through three machines, and older court filings with stamps over text produce errors. The error rate on clean digital PDFs and the error rate on a faxed handwritten chart are not in the same category, and vendors rarely demo the second.

Anything requiring legal judgment does not delegate. Whether a document is privileged, whether an objection is well founded, whether a fact pattern supports a claim, what to concede in negotiation. Software can propose a first pass at privilege review; it cannot be the reviewer of record.

Novel or non-standard documents break pattern-based work. Automation handles the eighty percent of matters that look like every other matter. The unusual scheduling order, the pro se filing, the settlement agreement with bespoke terms: these need a human immediately, and the system should be designed to route them out rather than attempt them.

Cite-checking and legal research summarisation still require verification of every citation. Do not build a workflow where a generated summary reaches a filing without an attorney confirming the underlying authority exists and says what the summary claims.

Client-facing text that touches substance is risky. An automated status update saying records were received is fine. An automated message that characterises case strength, timeline to resolution, or likely value is not fine, and no disclaimer fixes that.

How supervision should actually work

Treat an ai paralegal exactly as you would treat a new hire on their first month: give it narrow tasks, check the output, and expand scope only where the checking stops finding problems.

Assign a named supervisor per workflow. Not the firm generally. One person who owns the records intake automation and one who owns intake follow-up. That person reviews the exception queue daily and reviews a sample of successful runs weekly.

Design confidence thresholds into the workflow rather than bolting review on afterwards. Below a set confidence, the item goes to a human queue with the reasoning visible. Above it, the item proceeds but is logged. The log matters: you need to be able to reconstruct what the system did on any given matter six months later.

Keep humans in the loop at the points of irreversibility. Anything that files, sends externally, commits a deadline, or touches a client should require a click from a person. Anything internal and reversible can run unattended.

Review the exception rate as a metric. If a workflow is routing thirty percent of items to human review, it is not saving what you think it is saving and the underlying logic needs work. If it is routing under two percent, check whether it is being appropriately cautious or quietly failing.

Write down the boundary. A one-page document per workflow stating what the automation does, what it never does, and who reviews it. This is useful for training, for malpractice carriers, and for the moment a client asks how their file is handled.

What to do first

Pick one workflow, not five. Choose the one where the failure mode is visible and cheap: intake follow-up is usually the right starting point, because a missed message shows up immediately and the downside of a duplicate text is negligible.

Map it manually before automating it. Write out every step a human currently takes, including the ones nobody documented. Most firms discover at this stage that the process differs by practice group or by who happens to be handling it, and that inconsistency has to be resolved before software can help.

Then run the automation in shadow mode for two weeks. Let it produce output that a human compares against what actually happened, without acting on it. You will find the edge cases faster this way than by launching and correcting in production.

Alphovia builds this kind of operational automation for law firms, and the pattern above is roughly how we sequence it. But the sequencing matters more than the vendor. One workflow, documented, supervised, and measured, will teach you more about where an ai paralegal fits in your firm than any demo will.